The Signal
IonQ and EPB Report 95.8% Fewer Parameters in Grid AI Test
IonQ and Chattanooga utility EPB report that a hybrid quantum-classical model used 95.8% fewer trainable parameters than a classical neural network in a 33-bus grid test. Separate studies tested quantum-assisted approaches to distribution n
Quantum.world Team · Written with AI · fact-checked ·
A smaller Q-network, the same convergence mark
IonQ said on October 28 that a hybrid quantum-classical reinforcement-learning model developed with Chattanooga utility EPB reached a training benchmark with 95.8% fewer trainable parameters than its classical counterpart. In the comparison, a 16-qubit quantum model used 9,156 parameters; a classical network with 128 hidden units used 220,052. Both reached the study’s convergence threshold at episode 892.
The result concerns the size of a neural network used to choose grid-switching actions—not a 95.8% reduction in the number of physical grid-model parameters, grid equipment, or operating costs. The researchers tested the models on the IEEE 33-bus distribution system, using four random seeds for the model-size comparison. They report equal mean final reward for the 16-qubit and 128-unit models, while noting that the 16-qubit model did not find the exact optimum in every seed.
“Our results show that quantum-enhanced Q-networks can achieve higher rewards with fewer parameters than their classical counterpart, suggesting that such architectures are a promising direction for combinatorial power systems optimization.”
Daniel Germain et al., EPB Quantum and IonQ
Three approaches target distribution network reconfiguration
The work addresses distribution network reconfiguration, or DNR: selecting which switches to open or close to change power flows while keeping the network in a permitted configuration. Utilities use reconfiguration to pursue goals such as reducing line losses and balancing loads. The task becomes more complicated as grids add switching equipment and distributed resources such as renewables and microgrids.
The parameter-efficiency finding is one of three research efforts IonQ highlights. A second paper combines quantum sampling with a classical mixed-integer solver; its authors evaluated six networks ranging from 33 to 417 buses. For the largest test network, they report that hardware sampling found a configuration about 1.4% from the reference loss, while the guided classical solver reached a result within 1% of its target with lower median solver time than the unguided version across the six test systems.
A third paper uses a branch-exchange encoding with warm-start quantum imaginary-time evolution. Its authors report minimal-loss configurations in six of ten test cases and results 0.2% to 2.18% above the reported reference minimum in the other four. Those results come from distinct methods and test cases; they should not be read as a single head-to-head comparison.
Hardware runs test inference, not live grid control
For the reinforcement-learning study, the team also ran trained 12-qubit models on IonQ’s Forte Enterprise-1 processor. On the IEEE 33-bus case, the hardware-executed policy recovered the known optimal configuration. On the IEEE 69-bus case, its reported line loss was about 0.30% above the simulated optimum.
The paper’s 95.8% figure, however, comes from the 33-bus training comparison, not from the hardware demonstration. The hardware tests measured how trained models performed when making inference on benchmark systems; they do not show the system operating a utility’s live grid. The study also says its larger model-size comparison was not run on the 69-bus system because quantum circuit simulation was costly.
Preprints set the scope of the claim
The three papers appeared on arXiv in September 2026, with the two-stage DNR paper revised on October 12. The 95.8% result belongs to the study of quantum-enhanced deep reinforcement learning, whose authors describe the finding as a parameter-efficiency result for the tested architectures. Its benchmark and seed limits matter: the paper calls for testing larger configuration spaces to assess how the approach performs as the problem grows.
EPB and IonQ also frame the collaboration as a workforce effort. IonQ says EPB trained eight quantum fellows through a program supported by a $4 million National Institute of Standards and Technology grant awarded in 2024. For now, the concrete evidence is a set of benchmark studies, including inference runs on IonQ hardware—not deployment on EPB’s operating distribution grid.